[llvm] r216045 - InstCombine: Annotate sub with nuw when we prove it's safe

Hal Finkel hfinkel at anl.gov
Wed Aug 20 01:22:08 PDT 2014


----- Original Message -----
> From: "David Majnemer" <david.majnemer at gmail.com>
> To: llvm-commits at cs.uiuc.edu
> Sent: Wednesday, August 20, 2014 2:17:31 AM
> Subject: [llvm] r216045 - InstCombine: Annotate sub with nuw when we prove	it's safe
> 
> Author: majnemer
> Date: Wed Aug 20 02:17:31 2014
> New Revision: 216045
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=216045&view=rev
> Log:
> InstCombine: Annotate sub with nuw when we prove it's safe
> 
> We can prove that a 'sub' can be a 'sub nuw' if the left-hand side is
> negative and the right-hand side is non-negative.

I thought that we were specifically not doing this kind of thing because it can block further optimizations (as explained by Nick here: http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20140331/211474.html). Is the motivation to get around the depth limit in computeKnownBits? 

   -Hal

> 
> Modified:
>     llvm/trunk/lib/Transforms/InstCombine/InstCombine.h
>     llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp
>     llvm/trunk/test/Transforms/InstCombine/sub.ll
> 
> Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombine.h
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombine.h?rev=216045&r1=216044&r2=216045&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Transforms/InstCombine/InstCombine.h (original)
> +++ llvm/trunk/lib/Transforms/InstCombine/InstCombine.h Wed Aug 20
> 02:17:31 2014
> @@ -251,6 +251,7 @@ private:
>    bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
>    bool WillNotOverflowUnsignedAdd(Value *LHS, Value *RHS);
>    bool WillNotOverflowSignedSub(Value *LHS, Value *RHS);
> +  bool WillNotOverflowUnsignedSub(Value *LHS, Value *RHS);
>    Value *EmitGEPOffset(User *GEP);
>    Instruction *scalarizePHI(ExtractElementInst &EI, PHINode *PN);
>    Value *EvaluateInDifferentElementOrder(Value *V, ArrayRef<int>
>    Mask);
> 
> Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp?rev=216045&r1=216044&r2=216045&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp
> (original)
> +++ llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp Wed
> Aug 20 02:17:31 2014
> @@ -988,6 +988,20 @@ bool InstCombiner::WillNotOverflowSigned
>    return false;
>  }
>  
> +/// \brief Return true if we can prove that:
> +///    (sub LHS, RHS)  === (sub nuw LHS, RHS)
> +bool InstCombiner::WillNotOverflowUnsignedSub(Value *LHS, Value
> *RHS) {
> +  // If the LHS is negative and the RHS is non-negative, no unsigned
> wrap.
> +  bool LHSKnownNonNegative, LHSKnownNegative;
> +  bool RHSKnownNonNegative, RHSKnownNegative;
> +  ComputeSignBit(LHS, LHSKnownNonNegative, LHSKnownNegative, DL, 0);
> +  ComputeSignBit(RHS, RHSKnownNonNegative, RHSKnownNegative, DL, 0);
> +  if (LHSKnownNegative && RHSKnownNonNegative)
> +    return true;
> +
> +  return false;
> +}
> +
>  // Checks if any operand is negative and we can convert add to sub.
>  // This function checks for following negative patterns
>  //   ADD(XOR(OR(Z, NOT(C)), C)), 1) == NEG(AND(Z, C))
> @@ -1660,6 +1674,10 @@ Instruction *InstCombiner::visitSub(Bina
>      Changed = true;
>      I.setHasNoSignedWrap(true);
>    }
> +  if (!I.hasNoUnsignedWrap() && WillNotOverflowUnsignedSub(Op0,
> Op1)) {
> +    Changed = true;
> +    I.setHasNoUnsignedWrap(true);
> +  }
>  
>    return Changed ? &I : nullptr;
>  }
> 
> Modified: llvm/trunk/test/Transforms/InstCombine/sub.ll
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/sub.ll?rev=216045&r1=216044&r2=216045&view=diff
> ==============================================================================
> --- llvm/trunk/test/Transforms/InstCombine/sub.ll (original)
> +++ llvm/trunk/test/Transforms/InstCombine/sub.ll Wed Aug 20 02:17:31
> 2014
> @@ -506,6 +506,18 @@ define i4 @test42(i4 %x, i4 %y) {
>  ; CHECK-LABEL: @test42(
>  ; CHECK-NEXT: [[AND:%.*]] = and i4 %y, 7
>  ; CHECK-NEXT: [[AND1:%.*]] = and i4 %x, 7
> -; CHECK-NEXT: [[RET:%.*]] = sub nsw i4 %a, %b
> +; CHECK-NEXT: [[RET:%.*]] = sub nsw i4 [[AND]], [[AND1]]
> +; CHECK: ret i4 [[RET]]
> +}
> +
> +define i4 @test43(i4 %x, i4 %y) {
> +  %a = or i4 %x, -8
> +  %b = and i4 %y, 7
> +  %c = sub i4 %a, %b
> +  ret i4 %c
> +; CHECK-LABEL: @test43(
> +; CHECK-NEXT: [[OR:%.*]] = or i4 %x, -8
> +; CHECK-NEXT: [[AND:%.*]] = and i4 %y, 7
> +; CHECK-NEXT: [[RET:%.*]] = sub nuw i4 [[OR]], [[AND]]
>  ; CHECK: ret i4 [[RET]]
>  }
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory



More information about the llvm-commits mailing list